The Agent Skills Directory

[COMMAND_EXECUTION]: The skill constructs shell commands by directly interpolating variables (PROJECT, REMOTE, dir, slug) into strings that are then executed (e.g., rm -rf ${dir}, cp -r ${PROJECT} ${dir}/proj). The absence of shell escaping or sanitization on these variables makes the harness vulnerable to command injection if the arguments contain shell metacharacters.
[REMOTE_CODE_EXECUTION]: The skill facilitates remote execution by using SSH to run aiken check and other shell utilities on a user-provided remote host. The AI agent generates the code to be executed remotely, creating a path for code execution on external infrastructure.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data by reading Aiken source code files (validators/*.ak, lib/**/*.ak) from the local environment. This content is used to inform the agent's attack strategies and code generation. Maliciously crafted comments or code within the analyzed files could attempt to influence the agent's behavior or subvert the security audit process.
[DATA_EXPOSURE]: The workflow performs recursive copy operations (cp -r) of the specified project directory to local and remote scratch paths (/tmp/aiken-redteam). If the project argument is manipulated to point to sensitive system directories (e.g., ~/.ssh), the skill will attempt to copy and potentially exfiltrate that data to the remote build host via SCP.

aiken-validator-redteam